Dual denoising contrastive learning with multi-interest fusion for sequential recommendation

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Sequential recommendation models user behavior to predict future interactions.
• Multi-interest learning enhances recommendations by capturing diverse user preferences.
• Existing methods often overlook temporal preference shifts and noisy interaction data.

Purpose of the Study:

• To address limitations in sequential recommendation, specifically neglecting temporal dynamics and handling noisy data.
• To propose a novel dual denoising contrastive learning with multi-interest fusion approach (D2MFRec).
• To improve the accuracy and robustness of next-item prediction in sequential recommendation systems.

Main Methods:

• Developed a multi-interest aggregation module integrating single and multiple interest representations for comprehensive user embeddings.
• Introduced a dual denoising module to mitigate the impact of noisy interactions on user interest modeling.
• Employed a gated fusion mechanism for adaptive combination of diverse interest signals into a unified user representation.

Main Results:

• D2MFRec demonstrated superior performance compared to state-of-the-art baselines across three benchmark datasets.
• The proposed dual denoising and multi-interest fusion strategies significantly enhanced recommendation accuracy.
• The method effectively captured temporal user preference evolution and reduced noise impact.

Conclusions:

• D2MFRec offers a robust and effective solution for sequential recommendation by addressing temporal dynamics and data noise.
• The integration of multi-interest learning with denoising techniques provides a significant advancement in the field.
• The findings validate the effectiveness of the proposed approach for improving next-item prediction accuracy.